Re: [PATCH 1/2] drm/amdgpu: add debugfs for reset registers list

["Sharma, Shashank" <shashank.sharma@amd.com>]

 >> User only update the list of reg offsets on init, there is no
 >> predefined reg offset from kernel code.

I missed this comment in the last patch, and this makes me a bit 
confused. During the design phase, did we agree to have this whole list 
loaded from user ? which means that if user doesn't define the list at 
init, we will not send the trace_event().

Or was it kernel has a list, and user can modify if he wants to, and we 
will dump the values as per the register list.

@Christian ?

Regards
Shashank
On 2/8/2022 3:18 PM, Sharma, Shashank wrote:
> 
> 
> On 2/8/2022 2:39 PM, Somalapuram, Amaranath wrote:
>>
>>
>> On 2/8/2022 4:43 PM, Sharma, Shashank wrote:
>>> I thought we spoke and agreed about:
>>> - Not doing dynamic memory allocation during a reset call,
>> as there is a redesign debugfs call will happen during the application 
>> initialization and not during reset.
>>> - Not doing string operations, but just dumping register values by 
>>> index.
>> I think your referring to the second patch which happens during reset 
>> and no string operation in second patch.
> 
> Pls see my comment in the end.
> 
>>> NACK !
>>>
>>> - Shashank
>>>
>>> Amar,
>>> Apart from the long comment,there are a few more bugs in the patch, 
>>> which I have mentioned here inline. Please check them out.
>>>
>>> - Shashank
>>>
>>> On 2/8/2022 9:18 AM, Christian König wrote:
>>>> Am 08.02.22 um 09:16 schrieb Somalapuram Amaranath:
>>>>> List of register to be populated for dump collection during the GPU 
>>>>> reset.
>>>>>
>>>>> Signed-off-by: Somalapuram Amaranath <Amaranath.Somalapuram@amd.com>
>>>>> ---
>>>>>   drivers/gpu/drm/amd/amdgpu/amdgpu.h         |  3 ++
>>>>>   drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c | 60 
>>>>> +++++++++++++++++++++
>>>>>   2 files changed, 63 insertions(+)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h 
>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>>>> index b85b67a88a3d..78fa46f959c5 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>>>> @@ -1097,6 +1097,9 @@ struct amdgpu_device {
>>>>>       struct amdgpu_reset_control     *reset_cntl;
>>>>>       uint32_t ip_versions[HW_ID_MAX][HWIP_MAX_INSTANCE];
>>>>> +
>>>>> +    /* reset dump register */
>>>>> +    long            reset_dump_reg_list[128];
>>>>
>>>> I don't have time for a full review, but using long here certainly 
>>>> makes no sense.
>>>>
>>>> long is either 32bit or 64bit depending on the CPU architecture.
>>>>
>>>> Regards,
>>>> Christian.
>>>>
>> will change uint32_t.
>>>>>   };
>>>>>   static inline struct amdgpu_device *drm_to_adev(struct drm_device 
>>>>> *ddev)
>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c 
>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
>>>>> index 164d6a9e9fbb..dad268e8a81a 100644
>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_debugfs.c
>>>>> @@ -1609,6 +1609,64 @@ DEFINE_DEBUGFS_ATTRIBUTE(fops_ib_preempt, NULL,
>>>>>   DEFINE_DEBUGFS_ATTRIBUTE(fops_sclk_set, NULL,
>>>>>               amdgpu_debugfs_sclk_set, "%llu\n");
>>>>> +static ssize_t amdgpu_reset_dump_register_list_read(struct file *f,
>>>>> +                char __user *buf, size_t size, loff_t *pos)
>>>>> +{
>>>>> +    struct amdgpu_device *adev = (struct amdgpu_device 
>>>>> *)file_inode(f)->i_private;
>>>>> +    char *reg_offset;
>>>>> +    int i, r, len;
>>>>> +
>>>>> +    reg_offset = kmalloc(2048, GFP_KERNEL);
> 
> We also want to understand how does the value 2048 came into picture, 
> probably a macro which calculates the size preprocessing time will work 
> better.
> 
> #define #define N_REGS_DUMP_GPU_RESET 10
> #define BUFFER_SZ(N_REGS_DUMP_GPU_RESET * (sizeof uint64_t) + 1)
> 
> This first macro can be used later for the loop count for registers as 
> well.
> 
>>>>> +    memset(reg_offset,  0, 2048);
>>>>> +    for (i = 0; adev->reset_dump_reg_list[i] != 0; i++)
>>>
>>> This loop termination condition is incorrect, why are we running the 
>>> loop until adev->reset_dump_reg_list[i] != 0 ?
>>>
>>> What if I have 10 registers to dump, but my 4th register value is 0 ? 
>>> It will break the loop at 4 and we will not get all values.
>>>
>> agreed, i try to avoid one more variable in adev
> 
> Not by the cost of logic of course :).
> 
> Now you can run this loop here.
> 
> for (i = 0; i < N_REGS...; i++) {
>      register_value_copy_here;
> }
> 
>>>>> +        sprintf(reg_offset + strlen(reg_offset), "0x%lx ", 
>>>>> adev->reset_dump_reg_list[i]);
>>>>> +
> 
>>>>> +    sprintf(reg_offset + strlen(reg_offset), "\n");
>>>>> +    len = strlen(reg_offset);
>>>>> +
>>>>> +    if (*pos >=  len)
>>>>> +        return 0;
>>>>> +
>>>>> +    r = copy_to_user(buf, reg_offset, len);
>>>>> +    *pos += len - r;
>>>>> +    kfree(reg_offset);
>>>
>>> Also, why are we doing a dynamic memory allocation for reg_offest ? 
>>> We can simply use adev->reset_dump_reg_list[i] isnt't it ?
>>>
>>> simply
>>> for (i=0; i<num_of_regs;i++) {
>>>     copy_to_user(buf, adev->reg_list[i], sizeof(uint64_t));
>>> }
>>>
>>> Or without even a loop, simply:
>>> copy_to_user(buf, &adev->reg_list, num_regs * sizeof(uint64_t));
>>>
>>> - Shashank
>>
>> it will not be in user readable format for debugfs, (if non readable 
>> is acceptable I will change this)
>>
> 
> We are just adding 0x in front of the reg value, so honestly I don't see 
> a huge improvement in the user readability, but if you still want to do 
> the dynamic allocation of memory, add the register offset or name as 
> well, I mean then it should read like:
> 
> 0x1234 = 0xABCD
> 0x1238 = 0xFFFF
> 
> - Shashank
> 
>>> +
>>>>>
>>>>> +    return len - r;
>>>>> +}
>>>>> +
>>>>> +static ssize_t amdgpu_reset_dump_register_list_write(struct file 
>>>>> *f, const char __user *buf,
>>>>> +        size_t size, loff_t *pos)
>>>>> +{
>>>>> +    struct amdgpu_device *adev = (struct amdgpu_device 
>>>>> *)file_inode(f)->i_private;
>>>>> +    char *reg_offset, *reg;
>>>>> +    int ret, i = 0;
>>>>> +
>>>>> +    reg_offset = kmalloc(size, GFP_KERNEL);
>>>>> +    memset(reg_offset,  0, size);
>>>>> +    ret = copy_from_user(reg_offset, buf, size);
>>>>> +
>>>
>>> We are not allowing user to write into the list, so this whole 
>>> function can just be a NOOP.
>>>
>>> - Shashank
>> User only update the list of reg offsets on init, there is no 
>> predefined reg offset from kernel code.
>>>
>>>>> +    if (ret)
>>>>> +        return -EFAULT;
>>>>> +
>>>>> +    while ((reg = strsep(&reg_offset, " ")) != NULL) {
>>>>> +        ret  = kstrtol(reg, 16, &adev->reset_dump_reg_list[i]);
>>>>> +        if (ret)
>>>>> +            return -EINVAL;
>>>>> +        i++;
>>>>> +    }
>>>>> +
>>>>> +    kfree(reg_offset);
>>>>> +
>>>>> +    return size;
>>>>> +}
>>>>> +
>>>>> +static const struct file_operations 
>>>>> amdgpu_reset_dump_register_list = {
>>>>> +    .owner = THIS_MODULE,
>>>>> +    .read = amdgpu_reset_dump_register_list_read,
>>>>> +    .write = amdgpu_reset_dump_register_list_write,
>>>>> +    .llseek = default_llseek
>>>>> +};
>>>>> +
>>>>>   int amdgpu_debugfs_init(struct amdgpu_device *adev)
>>>>>   {
>>>>>       struct dentry *root = adev_to_drm(adev)->primary->debugfs_root;
>>>>> @@ -1672,6 +1730,8 @@ int amdgpu_debugfs_init(struct amdgpu_device 
>>>>> *adev)
>>>>>                   &amdgpu_debugfs_test_ib_fops);
>>>>>       debugfs_create_file("amdgpu_vm_info", 0444, root, adev,
>>>>>                   &amdgpu_debugfs_vm_info_fops);
>>>>> +    debugfs_create_file("amdgpu_reset_dump_register_list", 0644, 
>>>>> root, adev,
>>>>> +                &amdgpu_reset_dump_register_list);
>>>>>       adev->debugfs_vbios_blob.data = adev->bios;
>>>>>       adev->debugfs_vbios_blob.size = adev->bios_size;
>>>>